EGI-InSPIRE EGI-InSPIRE RI DDM solutions for disk space resource optimization Fernando H. Barreiro Megino (CERN-IT Experiment Support) on behalf of ATLAS DDM WLCG Collaboration Workshop 8 July 2010

EGI-InSPIRE RI Optimize space resources –Minimize unused space in sites –Keep sites full with interesting data –Improve control and monitoring Proposed solutions –DDM Accounting –Data ranking –Automatic replica reduction –Automatic DDM site exclusion Motivation

EGI-InSPIRE RI DDM Accounting

EGI-InSPIRE RI Disk space and usage information at sites collected by different agents and presented in a web frontend –Volumes of space, number of datasets and files registered at sites according to DDM (Vincent Garonne) possibility to break down volumes by metadata –Total and used disk space according to SRM DDM Accounting

EGI-InSPIRE RI Original main purpose: Follow up consistency between data volumes known by DDM and SRM DDM Accounting Good example of consistency: FZK running regular checks (Cedric Serfon)

EGI-InSPIRE RI Federation reports Other federations deploy more than they have pledged Space deployed by federations Distribution of resources amongst spacetokens Space pledged per federation Some federations don’t deploy their pledge at once (IFAE, IFIC-LCG2, UAM) (PIC) (LIP-COIMBRA, LIP- LISBON, NCG- INGRID-PT)

EGI-InSPIRE RI Management resource overviews Plot by Vincent Garonne

EGI-InSPIRE RI Physics groups reports Usage very different between groups Group quota converging to 25TiB booked per site

EGI-InSPIRE RI Data ranking

EGI-InSPIRE RI Monitoring of data ranking evolution Distinguish Computing Model replicas from additional replicas Custodial: Master replica (Must be kept) Primary: Following Computing Model (CM) Secondary: In excess of CM, may be deleted Default: Not ranked yet Introduction of the custodiality concept Start of meaningful ranking and bootstrapping Bootstrapping complete Space that could be recovered

EGI-InSPIRE RI Data ranking Globally it seems we have significant unused and recoverable space, but this depends on the site

EGI-InSPIRE RI Automatic replica reduction

EGI-InSPIRE RI Motivation Estimated pledge for T1 DATADISK until June 2011: 13PB Motivation

EGI-InSPIRE RI Reminder on DDM Popularity Angelos Molfetas (CERN) Only actions by official tools can be considered Try it out! Popularity.cern.ch Try it out! Popularity.cern.ch

EGI-InSPIRE RI Automatic replica reduction in theory Andrii Tykhonov (Jozef Stefan Institute, Ljubljana, Slovenia ) Agent that looks once a day for full MCDISK and DATADISK endpoints –Full endpoint: free space<min(10%, 25.0 TB) For each full site it will try to select secondary replicas to delete –Successfully cleaned endpoint: free space>min(20%, 50.0 TB) Selection of replicas for deletion based on DDM popularity –Older than 15 days –From oldest to newest –Not accessed recently Datasets will be moved into the deletion queue after a grace period of 24 hours. Evolution of the cleaning strategy from previous presentations

EGI-InSPIRE RI Automatic replica reduction in practice Some sites can not be completely cleaned: The data is too fresh and/or too popular. Actual deletion started this week. Until now our ops team was manually deleting the published lists

EGI-InSPIRE RI Somes notes on automatic replica reduction Early stage of experience with the system –Reduce secondary replicas more aggressively –Follow closely un-cleaned DATADISKs and MCDISKs –Detect exceptions and provide fine tuning by sites Sites who deploy disk space at once will be the ones to fully profit of this model Replica reduction is allied with the Dynamic Data Placement (D2PD) model

EGI-InSPIRE RI Future work on automatic replica reduction Last week’s decision: expand automatic replica reduction to SCRATCHDISK –SCRATCHDISK stores analysis output for users –At the moment: Fixed policy - datasets older than 30 days are automatically deleted –In the future: More flexible policy Keep between 20 and 10% free in spacetoken Once spacetoken has less than 10% free, remove data from oldest to newest until 20% free Data ranking or popularity do not play a role for this spacetoken

EGI-InSPIRE RI Automatic DDM site exclusion

EGI-InSPIRE RI Automatic DDM site exclusion Published GOCDB and OIM downtimes are taken into account automatically Full spacetokens will be excluded as destination Write Read All the information is propagated to the Site Status Board

EGI-InSPIRE RI Effort from ATLAS Distributed Computing in optimizing space resources provided by sites Need to become more aggressive in secondary data deletion Reduce need of manual operations Monitoring and deletion tools are in place and will be improved and tuned with experience Conclusions

EGI-InSPIRE RI Acknowledgements Simone Campana Alessandro Di Girolamo Stephane Jézéquel I. Ueda

EGI-InSPIRE RI Backup slides

EGI-InSPIRE RI Slide from Kors Bos Computing model

EGI-InSPIRE RI PanDA Dynamic Data Placement (D2PD) Slide from Kaushik De